Researchers at the Korea Advanced Institute of Science and Technology (KAIST) have published the first systematic analysis of the computational cost and energy consumption of AI agents. The study, presented at the 32nd IEEE International Symposium on High-Performance Computer Architecture (HPCA) in February 2026, found that these autonomous systems can consume up to 136.5 times more energy per query than conventional generative AI. The work signals that competitiveness in AI is broadening beyond model performance to encompass data-center and power infrastructure efficiency.
Repeated model calls drive the surge
AI agents differ from simple chatbots because they plan, use external tools like web search or code execution, and chain multiple reasoning steps. The KAIST team, led by Professor Minsoo Rhu of the School of Electrical Engineering, showed that this behavior leads to far more large language model (LLM) invocations than traditional chain-of-thought reasoning. Each time the agent calls an LLM to process a new judgment or response, the computational load multiplies.
The impact on hardware is stark. Response latency can increase by up to 153.7 times, and GPUs sit idle for as much as 54.5% of total execution time while external tools complete their tasks. This creates a new form of inefficiency: expensive high-performance chips remain underutilized even as the overall workload balloons.
Power demand at search scale
At a data-center level, an AI agent built on a 70-billion-parameter LLM - a scale comparable to current commercial services - consumed an average of 348.41 watt-hours per query. By contrast, a conventional generative AI system answering a straightforward question uses a fraction of that energy. The team then projected a scenario where 13.7 billion AI agent requests are generated daily, matching current Google search traffic.
Under that load, data-center power demand would reach approximately 198.9 gigawatts. That figure far exceeds the capacities of AI data centers under development today, which typically reach a few gigawatts, and equals roughly half the average power consumption of the United States.
"This study is the first to quantitatively show not only how AI is becoming more intelligent, but also how much electricity and cost are required to implement and sustain that intelligence," said Professor Minsoo Rhu. He added that as AI agents become widespread, "it will become increasingly important to take an integrated co-design approach that optimizes not only AI data-center infrastructure, but also AI agent models and power infrastructure."
Why this matters for science and research
The findings make one point concrete: scaling AI agent adoption forces trade-offs between capability and infrastructure cost that research labs and institutions cannot ignore. Scientists planning to deploy agents for literature review, experiment design, or collaborative reasoning will need to budget for power consumption equivalent to small industrial facilities. The open-source benchmarks and agent implementations released by the KAIST team provide a direct way for other researchers to measure and compare these hidden costs, making efficiency a quantifiable dimension in future AI system design.
Your membership also unlocks:
